Minutes_2023_01_10
Siu Kwan Lam edited this page Jan 10, 2023
·
1 revision
Attendees: Siu Kwan Lam, Andre Masella, Adrian Seyboldt, brandon willard, Graham Markall, Kaustubh Chaudhari, Luk, Matthew Murray, Remi, stuart, Todd A. Anderson, Guilherme Leobas, Larry Dong FPOC (last week): Andre FPOC (incoming): Siu
NOTE: All communication is subject to the Numba Code of Conduct.
Please refer to this calendar for the next meeting date.
- (Discussed in triage) https://numba.discourse.group/t/compilation-pipeline-compile-time-and-vectorization/1716
- May be lazy compilation in OrcJIT can help in cases where a lot of intermediate functions even without split-typing
- Adrian case:
- inlining of SIMD-optimizated may hide SIMD opportunity from caller context because LLVM SIMD pass disable future SIMD-optimization on code it touched.
- Graham has a OrcJIT branch that will be ready in weeks for us to test
-
https://github.com/numba/numba/pull/7270 - Add support for numpy.moveaxis/normalize_axis_index and normalize_axis_list in nopython mode
- Assigned Graham to follow up on it
- Numba FFT support via SciPy (#5864)
- Decided to try ripping pointers using the same technique as getting BLAS symbols from NumPy since we are already trying to reduce SciPy dependency.
-
#8442, need to decide what happens in the case of
__repr__
when there's no definition in compiled code, at present, it calls objmode and warns that it's doing that.- Go with nopython
__repr__
even if the resulting string format will be different.
- Go with nopython
-
https://github.com/numba/numba/pull/7009
- Q: Should we require the use of name arguments instead of positional index?
- A: Yes
- Stuart RFC announcement: numba#8700
- Brandon reminds us of https://github.com/numba/numba-scipy/pull/92 - Add basic CSC and CSR sparse matrix support
- numba#8668 - Unexpected output type on Windows when calling f.types on ufunc
- numba#8669 - Convert static PyTypeObject structures to heap allocated types.
- numba#8676 - numba dict converts between int types
- numba#8678 - Remaining py3.11 with-lifting failures
- numba#8679 - njit(cache=True) is not idempotent (with generators?)
- numba#8681 - Numba crashes Python with simple function containing numpy.exp
- numba#8682 - Privatization of variables defined inside prange
- numba#8686 - Numba compiled functions releasing the GIL often fail to scale in parallel when Numpy and/or lists are involved
- numba#8687 - No definition for lowering <built-in method twoD_impl of _dynfunc._Closure object at 0x2b60ce355458>(array(float32, 2d, A), omitted(default=None)) -> float64
- numba#8693 - Access to outer variables in nested functions
- numba#8694 - pytest doctest can't discover functions with @vectorize decorator
- numba#8698 - NumPy 1.25 support
- numba#8700 - RFC: Separate NumPy and CPython implementions/functionality.
- llvmlite#897 - Support LLVM 15
- numba#8667 - Create Dockerfile to allow easy access/testing of Numba with Python 3.11 support
- numba#8672 - add support for dict types in np.asarray
-
numba#8683 - Incorrect result when implicitly broadcasting with
parallel=True
- numba#8684 - Can we add the jitclass into the signature
- numba#8689 - Memory leak when called function raises error
- numba#8692 - building a code that utilizes Numba and mpi4py
- numba#8695 - Unable to install
- numba#8670 - [WIP] Python 3.11 tracing
- numba#8674 - Enforce that no warnings are raised on listing of tests.
-
numba#8675 - Make
always_run
test decorator a tag and improve shard tests. - numba#8677 - Add support for min and max on boolean types.
- numba#8680 - Adjust flake8 config to be compatible with flake8=6.0.0
- numba#8685 - Implement repr for numba types
- numba#8688 - Implement np.nanargmin/nanargmax
- numba#8690 - [CI only] Testing #8620 with NumPy versions < 1.24
- numba#8691 - NumPy 1.24 (PR for review)
- numba#8696 - Don't document AArch64 installation separately
- numba#8697 - Close stale issues after 7 days
- numba#8671 - Debugging #8544
- merged - numba#8673 - Enable the CUDA simulator tests on Windows builds in Azure CI.